onnx手动操作

• 使用onnx.helper可以进行onnx的制造组装操作：

对象	描述
ValueInfoProto 对象	张量名、张量的基本数据类型、张量形状
算子节点信息 NodeProto	算子名称(可选)、算子类型、输入和输出列表(列表元素为数值元素)
GraphProto对象	用张量节点和算子节点组成的计算图对象
ModelProto对象	GraphProto封装后的对象

方法	描述
onnx.helper.make_tensor_value_info	制作ValueInfoProto对象
onnx.helper.make_tensor	使用指定的参数制作一个张量原型(与ValueInfoProto相比可以设置具体值)
onnx.helper.make_node	构建一个节点原型NodeProto对象 (输入列表为之前定义的名称)
onnx.helper.make_graph	构造图原型GraphProto对象（输入列表为之前定义的对象）
make_model（graph， **kwargs）	GraphProto封装后为ModelProto对象
make_sequence	使用指定的值参数创建序列
make_operatorsetid
make_opsetid
make_model_gen_version	推断模型IR_VERSION的make_model扩展，如果未指定，则使用尽力而为的基础。
set_model_props
set_model_props
make_map	使用指定的键值对参数创建 Map
make_attribute
get_attribute_value
make_empty_tensor_value_info
make_sparse_tensor

提取出一个子模型

import onnx  onnx.utils.extract_model('whole_model.onnx', 'partial_model.onnx', ['22'], ['28']) 

提取时添加额外输出

onnx.utils.extract_model('whole_model.onnx', 'submodel_1.onnx', ['22'], ['27', '31'])  # 本来只有31节点输出，现在让27节点的值也输出出来

使用（尝试构建一个模型）

import onnx 
from onnx import helper 
from onnx import TensorProto 
import numpy as npdef create_initializer_tensor(name: str,tensor_array: np.ndarray,data_type: onnx.TensorProto = onnx.TensorProto.FLOAT
) -> onnx.TensorProto:# (TensorProto)initializer_tensor = onnx.helper.make_tensor(name=name,data_type=data_type,dims=tensor_array.shape,vals=tensor_array.flatten().tolist())return initializer_tensor# input and output 
a = helper.make_tensor_value_info('a', TensorProto.FLOAT, [None,3,10,  10]) 
x = helper.make_tensor_value_info('weight', TensorProto.FLOAT, [10, 10]) b = helper.make_tensor_value_info('b', TensorProto.FLOAT, [None,3, 10,10]) 
output = helper.make_tensor_value_info('output', TensorProto.FLOAT, [None,None,None, None]) # Mul 
mul = helper.make_node('Mul', ['a', 'weight'], ['c']) # Add 
add = helper.make_node('Add', ['c', 'b'], ['output_of_liner']) # Conv
conv1_W_initializer_tensor_name = "Conv1_W"
conv1_W_initializer_tensor = create_initializer_tensor(name=conv1_W_initializer_tensor_name,tensor_array=np.ones(shape=(1, 3,*(2,2))).astype(np.float32),data_type=onnx.TensorProto.FLOAT)
conv1_B_initializer_tensor_name = "Conv1_B"
conv1_B_initializer_tensor = create_initializer_tensor(name=conv1_B_initializer_tensor_name,tensor_array=np.ones(shape=(1)).astype(np.float32),data_type=onnx.TensorProto.FLOAT)conv_node = onnx.helper.make_node(name="Convnodename",  # Name is optional.op_type="Conv",       # Must follow the order of input and output definitions. # https://github.com/onnx/onnx/blob/rel-1.9.0/docs/Operators.md#inputs-2---3inputs=[ 'output_of_liner', conv1_W_initializer_tensor_name,conv1_B_initializer_tensor_name ],outputs=["output"],kernel_shape= (2, 2), #pads=(1, 1, 1, 1),
)# graph and model 
graph = helper.make_graph([mul, add,conv_node], 'test', [a, x, b], [output],initializer=[conv1_W_initializer_tensor, conv1_B_initializer_tensor,],) 
model = helper.make_model(graph) # save model 
onnx.checker.check_model(model) 
print(model) 
onnx.save(model, 'test.onnx') ###################EEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEVVVVVVVVVVVVVVVVVVVVVVVVVVVVVAAAAAAAAAAAAAAAAAAAAAAAAALLLLLLLLLLLLLLLLLLLLLLLL#########
import onnxruntime 
# import numpy as np sess = onnxruntime.InferenceSession('test.onnx') 
a = np.random.rand(1,3,10, 10).astype(np.float32) 
b = np.random.rand(1,3,10, 10).astype(np.float32) 
x = np.random.rand(10, 10).astype(np.float32) output = sess.run(['output'], {'a': a, 'b': b, 'weight': x})[0] print(output)

• https://github.com/NVIDIA/TensorRT/tree/master/tools/onnx-graphsurgeon